Mohamed Elgaar

Mohamed Elgaar

PhD Student

I’m a Ph.D. candidate in Computer Science at the University of Massachusetts Lowell, advised by Professor Hadi Amiri, where I study data-efficient LLM training and training dynamics through linguistic complexity signals. In 2025, I was a Research Intern at Google DeepMind working on multilingual factuality evaluation for Gemini. Previously, I received my BS in Computer Science from KAIST.

My work connects two themes: using linguistic complexity to make training more efficient and stable, and enabling fine-grained linguistic control over model outputs.

  • Data-efficient LLM training: data ordering/scheduling, data curation and selection for large-scale pretraining.
  • Training dynamics & interpretability: scaling laws, learning phases, difficulty signals, capability evolution analysis.

Current directions

  • Pretraining curricula at scale: data ordering strategies for single-epoch pretraining; analyzing training dynamics and compute efficiency across model scales.
  • RL for controlled generation: reward shaping for constraint satisfaction, studying how data scheduling improves RL stability.

Selected Publications

Full publication list →

News

  • [Dec 2025] Released a preprint on curriculum learning for LLM pretraining (learning dynamics analysis).
  • [Nov 2025] Linguistically-Controlled Paraphrase Generation presented at EMNLP 2025.
  • [July 2025] MedDecXtract presented at ACL 2025 Demo Track.
  • [May 2025] Joined Google DeepMind as a Research Intern (Gemini multilingual factuality).
  • [Oct 2024] Released the P-Masking / LingGen preprint on multi-attribute controlled generation.
  • [Dec 2023] Presented Ling-CL at EMNLP 2023.